# Reproduce the missing-task sensitivity

This October 26, 2026 analysis uses the unchanged public allocation from the
October 8 coding study. It asks which missing task outcomes could change the
direction of the full 228-task descriptive difference. It adds no model runs,
grades, signature checks, policy verdicts or probability estimates.

Download [reproduce.py](reproduce.py) and the original
[allocated-outcomes.csv](../paired-coding-study/allocated-outcomes.csv).
Using Python 3.10 or later, run:

```sh
python3 reproduce.py --csv allocated-outcomes.csv
```

The script verifies the CSV SHA-256, keeps only primary attempts, preserves all
observed grades, and enumerates possible net differences for the eight attempted
tasks without complete paired grades. It then adds hypothetical net outcomes for
the 20 unattempted tasks. Compare the output with [sensitivity.json](sensitivity.json).
The two tasks held at setup are included among the 20 unattempted tasks. Repeats
do not add primary tasks and are excluded from this calculation.

Candidate-minus-baseline net means gains minus losses. A fully resolved allocation
uses its original category weights, which equal category task counts divided by
228. Its weighted percentage-point difference therefore equals 100 times the net
divided by 228. The observed 200-pair weighted estimate uses different within-category
denominators and must not be substituted into this calculation.

The original complete-pair estimate and its uncertainty interval are unchanged.
Scenario bounds are neither confidence intervals nor likely outcomes. They do
not assume missingness is random, establish superiority or equivalence, or make
the tested tasks representative of another workload. A later run would be a new
observation, not a recovery of an unobserved historical outcome.

Source identity: the CSV hash is
`9471edfbd90072a756a7928cf3c37db757d173c53498acf4b7d7d018a0d3302d`.
The file and its provenance are pinned in site revision
`b54bd2e883b84ac01616fc8b12a806b1eea2e655`. These public derivatives cannot
authenticate the private source archive.

- [Original allocation at the reviewed revision](https://github.com/invarlock/invarlock-site/blob/b54bd2e883b84ac01616fc8b12a806b1eea2e655/public/evidence/paired-coding-study/allocated-outcomes.csv)
- [Original numerical results](../paired-coding-study/results.json)
- [Original methods](../paired-coding-study/report.md)
- [Original provenance](../paired-coding-study/provenance.json)